A Pulse-Feature LSTM Framework with Temporal Variance-Based Prediction Filtering for rPPG-Based Blood Pressure Estimation

Non-contact blood pressure (BP) estimation from remote photoplethysmography (rPPG) can yield unstable window-level predictions. This study combines a pulse-feature long short-term memory (LSTM) estimator with temporal variance-based prediction filtering. Fifty-two features from the rPPG waveform and its first and second derivatives were arranged into sliding-window sequences for systolic BP (SBP) and diastolic BP (DBP) estimation. The filter selects locally stable predictions from temporally ordered prediction sequences without using ground-truth BP values. Performance was first assessed in 10 repeated subject-dependent randomized-block experiments. Filtering reduced the SBP mean absolute error (MAE) from 7.04±1.52 to 5.34±1.62 mmHg and the DBP MAE from 4.63±1.32 to 3.07±0.97 mmHg while retaining 76.39±7.59% of predictions. A subject-specific mean predictor was also competitive under the subject-dependent setting and achieved lower DBP MAE at full coverage than the filtered LSTM. Chronological rolling-origin and leave-one-subject-out (LOSO) evaluations further showed that the improvement persisted for later intervals from the same subjects but not for unseen subjects. Because the filter operates only on model outputs, it requires neither additional trainable parameters nor model retraining. These findings show a low-overhead accuracy–coverage trade-off for within-subject rPPG-based BP estimation, while subject-independent generalization remains limited.